DeepSeek-V4-Flash reaches 105 tok/s on two 4090D cards after Triton kernel rewrites
iSevenDays · reddit · 2026-07-24
A Reddit user reports running DeepSeek-V4-Flash at about 105 tok/s on two Nvidia 4090D 48G cards with vLLM, after re-implementing Blackwell-only kernels in Triton.
Key details:
- Recreated DeepGEMM, FlashInfer sparse-MLA, and block-scaled FP8 paths for sm89 GPUs.
- Uses a patched setup on a Dell R740 with P2P enabled.
- Claims 2–3x better throughput for parallel agentic workloads and around 262k context.
- Includes commands for both the vLLM setup and a fully packed llama.cpp run.
The post is essentially a field report on squeezing modern model performance out of non-Blackwell hardware.
More from Infra
- Nearly half of Claude revenue now flows through cloud marketplace commitments, up from 11% — BenBajarin · 2026-10-26
- Anthropic ARR goes vertical after price cuts: tokens up 4x, $100B+ ARR back in play — GavinSBaker · 2026-10-26
- Alex Banks argues power caps GPU supply, real risk is demand cooling first — thealexbanks · 2026-10-26
- Analyst: Ant has the most fungible AI workload of any model provider today — BenBajarin · 2026-10-26
- Anthropic S-1: training infra edge cut per-task cost 40% and doubled revenue growth — BenBajarin · 2026-10-26
- Rewriting scaling law study: optimal synthetic data share rises from 10% to 32% with compute — heghbalz · 2026-10-26